The problem

1.1Returns and fit uncertainty in online apparel

Apparel is the category where online shopping most often fails. Coresight Research found an average return rate of 24.4% for online apparel orders in the United States over the twelve months to March 2023, and 53% of the 100 apparel brand and retail decision-makers it surveyed named size and fit as the top reason [1]. Across all US retail, the National Retail Federation and Happy Returns expect USD 849.9 billion of merchandise to be returned in 2025, equal to 15.8% of annual sales, with online sales returned at a rate of 19.3% [8].

Every return carries reverse shipping, inspection, repackaging and markdown costs, and some returned apparel is never resold. The root cause is the shopper's uncertainty about how a garment will look on their own body. A studio photograph on a professional model cannot answer that question, in any market.

1.2Virtual try-on has become mainstream

Generative virtual try-on renders a garment onto a photograph of the shopper. It moved from research to mass-market products in about three years. Google launched try-on from a shopper's own photo in US Search in July 2025 and extended it to the United Kingdom and India in December 2025 [12, 2]. Inditex reports more than seven million sessions of Zara Try-On across 43 markets [13]. Alibaba's Tstars-Tryon runs inside the Taobao app and has served several million users [3]. Google Cloud now sells a dedicated try-on model to enterprises [14].

Shoppers in many markets are therefore learning to expect try-on. The open question is whether the technology works for everyone: for every garment shoppers actually buy, and for every body that wears it.

1.3Current models are built for a narrow wardrobe

Most try-on research treats a garment as a single stitched piece with a fixed shape, such as a T-shirt, a dress or a pair of trousers, photographed flat or on a studio model. VITON-HD, the standard benchmark, contains only upper-body garments [4]. Real catalogues and real shoppers are far more varied, in six ways that matter to a generative model.

Evidence of these gaps is uneven. Layering and in-the-wild photos now have dedicated papers, but garments outside the stitched Western wardrobe have almost none. BD-VITON, published in March 2026, is the only try-on benchmark for culturally specific clothing found in this review. It covers Bangladeshi sarees, panjabi and salwar kameez in 1,013 pairs, retrains three models released before 2024, and is not cleared for commercial use [6]. DIVA, a 2024 workshop paper, studied Indian try-on at 720×540 resolution [20], and IndoFashion offers 106,000 images of Indian ethnic wear for classification, with no try-on pairs [19]. A widely used curated list of try-on research, updated in September 2026, has no entry on draped or cultural garments [21]. Beyond these, this review found no try-on benchmark for the garment traditions of other regions.

1.4Representation and fairness are unmeasured

Try-on models change pixels on a person's body, so they can alter the person as well as the garment. SiCo reports that try-on training sets are dominated by slim bodies [22]. This review found no published fairness audit of try-on models by skin tone or body shape. Industry recognises the issue: when Google launched apparel try-on in 2023, it chose its models across the Monk Skin Tone scale and sizes XXS to 4XL [23, 24].

For shoppers everywhere the risks are concrete: skin tone drifting lighter, a body reshaped toward a slimmer norm, or the loss of personal and cultural details such as tattoos, jewellery, a headscarf or turban, mehndi or a bindi. None of these is captured by the metrics the field reports today, and a try-on product that serves a global market has to be measured against all of them.

1.5Licensing blocks the obvious path

A company cannot simply download the best research model and adapt it. VITON-HD is licensed CC BY-NC 4.0, and DressCode is not released to private companies [4, 5]. Leading open try-on checkpoints, including OOTDiffusion, IDM-VTON, CatVTON and FitDiT, are released under CC BY-NC-SA 4.0 [25, 26, 18, 27]. The newest strong open image editor, Qwen-Image-2.1, uses a research-only licence, and most FLUX models use a non-commercial licence that extends to fine-tuned derivatives [28, 29]. A commercial try-on model for a global market therefore needs its own data and a carefully chosen base model.

1.6FabricVTON's starting point

FabricVTON runs photo try-on as a managed GPU service with automatic garment-category detection, request batching and measured cost reporting [7]. This engineering is valuable, and it also shows clearly what is missing:

• No model trained on FabricVTON's own data yet. The production model is an open checkpoint that knows three garment categories: tops, bottoms and one-pieces [30]. Its behaviour on layered, multi-piece and draped garments, and across skin tones and body shapes, has not been measured.
• Resolution is limited. The current model generates at 576×864, and its authors list resolution and traces of the previous garment among its known limitations [30].
• Quality evidence is thin. Speed optimisations have so far been judged by eye on a few image pairs, which is not enough evidence for a research or product claim.
• No video or live capability of its own.The model draws one still image in about 6.5 seconds; live try-on needs a new frame every 33 to 66 milliseconds. FabricVTON's live page currently runs on a licensed third-party engine.

1.7Shoppers want to see garments move

A photo shows how a garment looks standing still. Shoppers also want to see how it moves, hangs and fits when they turn, walk or raise an arm, and on their own camera rather than in an uploaded photo. Live try-on has just become possible. Decart's live try-on model, Lucy VTON 3.5, redraws each camera frame at 720p, and Decart has raised more than USD 450 million [72, 73]. Google's Doppl app turned try-on photos into short videos before it closed in April 2026 and moved into Search, and FASHN, Luma and Runway now offer offline video try-on or wardrobe editing [74, 75, 76, 77].

The engines are closed and costly. The figure and table below compare one try-on on each path.

The one open research system for live video try-on, LiveVVT, reaches about 22 frames per second at 512×384, but it still needs a body mask and pose map for every frame and no code had been released as of September 2026 [71]. A commercial live model therefore has to be built, and built on data a company can legally use.

1.8Problem statement